Dear Pradeep,
I just realised that my answer may have been a bit terse. Let me
elaborate a litle: Finding the marginal effect of a variable in a
logistic regression is finding the first derivative with respect to
that x: {d pr(foreign=1)}/dx. Note that:
pr(foreign=1)= exp(xb)/(1+xb) , lets call that L(xb), and
xb=_b[-cons]+_b[weight]*weight+_b[price]*price+_b[mpg]*mpg+_b[mpg2]
*mpg^2
You can now apply the chain rule {d pr(foreign =1)}/dx = dL/(d xb) *
(d xb)/dx. It so happens that L(xb) is the cumulative probality
function of the logistic distribution. The deriviative of a
cumulative probability function is the probability density function
of that distribution. The probability density function of the
logistic is: exp(xb)/(1+exp(xb))^2. So:
dL/(d xb) = exp(xb)/(1+xb)^2
(d xb)/(d mpg) = _b[mpg] + 2*_b[mpg2]*mpg
evaluating this function at the mean for each variable gives you the
marginal effect of mpg when all variables have their mean value. Not
that you can thus calculate the marginal effect of weight, price and
mpg, but not of mpg^2, which makes substantive sense.
I hope that this clarifies my earlier post,
Maarten
--- In [email protected], "maartenbuis" <maartenbuis@y...>
wrote:
> Dear Pradeep,
>
> Calculating the marginal effect for a logistic regression when one
of
> the explanatory variables is also entered with an quadratic term
can
> be done by `hand' (well, by -nlcom- actually) but it gets a bit
ugly,
> see the example below.
>
> Notice that the first couple of lines of the -nlcom- is the
> probability density function of the logistic: exp(xb)/(1+exp(xb))
^2.
> The last lines of the -nlcom- command is the derivative of xb with
> respect to x of interest, in this cas mpg: _b[mpg]+2*_b[mpg2]*(mean
> mpg).
>
> Maarten
>
> #delim ;
> sysuse auto;
>
> gen mpg2=mpg^2;
>
> logit foreign price weight mpg mpg2;
>
> sum price, meanonly;
> local price = r(mean);
>
> sum weight, meanonly;
> local weight = r(mean);
>
> sum mpg, meanonly;
> local mpg = r(mean);
>
> nlcom(
> exp(_b[_cons] + _b[price]*`price' + _b[weight]*`weight' + _b
> [mpg]*`mpg' + _b[mpg2]*`mpg'^2)/
> (1+exp(_b[_cons] + _b[price]*`price' + _b[weight]*`weight' +
_b
> [mpg]*`mpg' + _b[mpg2]*`mpg'^2))^2*
> (_b[mpg]+2*_b[mpg2]*`mpg')
> );
>
> --- pradeep.kurukulasuriya@y...> wrote:
> > I have a logit model with a linear and quadratic variable along
> with some
> > other variables (call it stuff). I am trying to estimate the
> marginal
> > effects for different sets variables in my model where each set
is
> comprised
> > of a linear and quadratic term. Is there a ado file for this that
> does not
> > require the differentiation to be done manually (and then run a
> proceedure
> > like nlcom) or is that the best there is?
>
>
>
>
> *
> * For searches and help try:
> * http://www.stata.com/support/faqs/res/findit.html
> * http://www.stata.com/support/statalist/faq
> * http://www.ats.ucla.edu/stat/stata/
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
